fix cases where losing epoch get a huge negative loss value

[stash86]

Summary

On some cases, you can have losing epoch (total profit< 0), but it can have a big negative loss value, which make the bot think it's the best epoch. This should fix such bug

Solve the issue: #10192

Quick changelog

• <change log 1>
• <change log 1>

What's new?

[coveralls]

coverage: 94.666% (-0.01%) from 94.679%
when pulling 6306fcf on stash86:bt-metrics
into 19d6ce5 on freqtrade:develop.

[xmatthias]

I'm not sure it does if i'm honest.

i'm using the following test cases to look at this calculation - some results are from the issue - some are artificial to see which value has which influence ...

DRAWDOWN_MULT = 0.075

results = [
   (30.785, 0.35 / 100),
    (-221.333, 15.98 / 100),
    (-315.876, 40.04 / 100),
    (-459.923, 50.84 / 100),
    (-581.544, 68.98 / 100),
    (64.167, 5.31 / 100),
    (65.167, 5.31 / 100),
    (66.167, 5.31 / 100),
    (-628.566, 74.53 / 100),
    (602.105, 16.04 / 100),
    (602.105, 26.04 / 100),
    (602.105, 66.04 / 100),
    (602.105, 120.04 / 100),  # Drawdown after profit
    (-494.568, 70.59 / 100),
    (-251.262, 35.02 / 100),
    (-429.174, 46.65 / 100),
    (-759.062, 82.04 / 100),
]
for (total_profit, rel_drawdown) in results:

    res = -1 * (total_profit * (1 - rel_drawdown * DRAWDOWN_MULT))

    loss_value = total_profit * (1 - rel_drawdown * DRAWDOWN_MULT)

    if (total_profit < 0) and (loss_value > 0):
        pass
    else:
        loss_value = -1 * loss_value
    print(f"{res=:>10.7g}\t\t{loss_value=:>10.7g}\t\t{total_profit=:.3f},\t {rel_drawdown=:.2%} {res == loss_value}")

For explanation:
res is the current calculation
loss_value is the calculation in this branch

Output:

res= -30.77692		loss_value= -30.77692		total_profit=30.785,	 rel_drawdown=0.35% True
res=  218.6803		loss_value=  218.6803		total_profit=-221.333,	 rel_drawdown=15.98% True
res=  306.3902		loss_value=  306.3902		total_profit=-315.876,	 rel_drawdown=40.04% True
res=  442.3861		loss_value=  442.3861		total_profit=-459.923,	 rel_drawdown=50.84% True
res=  551.4578		loss_value=  551.4578		total_profit=-581.544,	 rel_drawdown=68.98% True
res= -63.91145		loss_value= -63.91145		total_profit=64.167,	 rel_drawdown=5.31% True
res= -64.90747		loss_value= -64.90747		total_profit=65.167,	 rel_drawdown=5.31% True
res= -65.90349		loss_value= -65.90349		total_profit=66.167,	 rel_drawdown=5.31% True
res=  593.4307		loss_value=  593.4307		total_profit=-628.566,	 rel_drawdown=74.53% True
res= -594.8617		loss_value= -594.8617		total_profit=602.105,	 rel_drawdown=16.04% True
res= -590.3459		loss_value= -590.3459		total_profit=602.105,	 rel_drawdown=26.04% True
res= -572.2827		loss_value= -572.2827		total_profit=602.105,	 rel_drawdown=66.04% True
res= -547.8975		loss_value= -547.8975		total_profit=602.105,	 rel_drawdown=120.04% True
res=  468.3843		loss_value=  468.3843		total_profit=-494.568,	 rel_drawdown=70.59% True
res=  244.6626		loss_value=  244.6626		total_profit=-251.262,	 rel_drawdown=35.02% True
res=  414.1583		loss_value=  414.1583		total_profit=-429.174,	 rel_drawdown=46.65% True
res=  712.3569		loss_value=  712.3569		total_profit=-759.062,	 rel_drawdown=82.04% True

The 2 don't deviate - not in ANY of the provided cases (not from the issue, nor in the artificial ones.

---

To be perfectly honest, i don't think the base calculation of this works the way it is now (the sign is just a small problem here).
drawdown is a value that's in the range of 0-1+ (120% can happen if you're in profit and then lose it all and more).
Profit however can be 0.1 (trading BTC?) - or 1000 (trading USDT).
The way it's combined leads to potentially big numbers (profit) which are either positive or negative - with VERY small numbers (we talk about numbers from 0-~1.5) - which are multiplied by 0.075 ... making a small number even smaller - reducing the impact of drawdown to almost nothing (visible in the test-cases with 602$ profit).

[xmatthias]

Closing this in favor of #10221 - which revises the calculation completely.